pm-claude-skills

What It Does

pm-claude-skills is a professional skill library for AI coding and general-purpose agents. It ships 466 SKILL.md files (version 46.0.0, MIT) covering product management, engineering, customer success, marketing, legal, finance, HR, sales, design, security, compliance, research, and more. Each skill is a structured markdown instruction file that teaches an agent how to produce a professional artifact: a PRD, an A/B test plan, an SOC2 readiness score, a sprint commitment, a competitor teardown, a pentest report. The library is consumed four ways: as native SKILL.md files installed into Claude, Hermes, Codex, or OpenClaw; as rule files exported for Cursor, Windsurf, and Aider; as an MCP server that exposes the library over stdio to any MCP client; and as an in-browser playground. The target users are professionals (PMs, engineers, managers, founders, marketers, lawyers) who want a deep catalog of ready-made skills that drop into whichever agent they already use.

The Good

The delivery plumbing is unusually complete and dependency-free. The installer (bin/cli.mjs) is pure Node standard library with no runtime dependencies. It supports eight targets in one command: native SKILL.md installs for Claude, Hermes, Codex, and OpenClaw (copying skill folders into the right home directory), and generated rule-file exports for Cursor (.mdc), Windsurf (.md), Aider (.md), and Kilocode (.md). A real install into a temp target copied all 466 skills with their SKILL.md, references/, and templates/ subdirectories intact, then printed a correct "Restart hermes" hint. Dry-run mode (--dry-run) previews every file path before writing. The MCP server (mcp/server.mjs) is also pure stdlib, speaks JSON-RPC 2.0 over stdio, and exposes six tools: list_skills, search_skills, get_skill, run_skill (which uses MCP sampling so generation runs on the client's own model with no server-side API key), list_workflows, and get_workflow. It ships proper server.json (the official MCP schema), glama.json, and smithery.yaml for registry indexing.

Every one of the 466 skills passes a structural validator, and the validator is part of the repo. scripts/skillcheck.mjs parses each SKILL.md, checks for valid frontmatter, requires a name and description, enforces a "Use when" trigger clause in the description, and flags skills missing required sections. Running it across all 466 skills returns 0 error(s) · 3 warning(s) · All skills valid. The 3 warnings are minor (three descriptions lack an explicit "Use when" trigger phrase). This means the catalog is internally consistent at scale, which is rare for a 466-item skill bundle. The same validator runs in CI (.github/workflows/skillcheck.yml).

32 skills ship real, runnable Python calculators that produce correct numbers. These are not stubs. skills/sprint-planning/scripts/capacity_calculator.py is a pure-stdlib argparse CLI that takes team size, sprint days, focus hours, historical velocity, availability, and carryover, and returns a computed sprint commitment. Running it with --team 5 --days 10 --velocity 30 --availability 0.8 --carryover 5 produces 240 available focus hours, a velocity-adjusted 24.0 points, a recommended 19.2-point commitment, and 14.2 points of capacity for new work. The catalog also includes runway-monte-carlo/scripts/runway_sim.py (a Monte Carlo cash-runway simulation), saas-metrics/scripts/saas_metrics.py, csat-nps-analysis/scripts/nps.py, roi-estimator/scripts/roi.py, soc2-readiness/scripts/soc2_score.py, gdpr-compliance/scripts/ropa_check.py, pricing-calculator/scripts/pricing.py, and unit-economics/scripts/unit_econ.py. All are pure standard library with no third-party dependencies and no network access, as the SECURITY.md explicitly guarantees.

The maturity tiers are real and backed by an eval harness. skill-tiers.json classifies 49 skills as production-ready, 12 as experimental, and the rest as stable. The production tier includes the headline skills (prd-template, rice-prioritisation, ab-test-planner, sprint-planning, go-to-market, churn-analysis, incident-postmortem, skill-security-auditor). The classification is not arbitrary: evals/run-evals.mjs is an LLM-as-judge harness that runs each skill on test cases and scores output 1-5 on four dimensions (structure, completeness, usefulness, grounding). Results render on a public Skill Leaderboard. The eval runner is cost-conscious by design (roughly $0.30 per full run, Sonnet judge, content-hash caching to skip unchanged skills, a --changed flag for per-PR evals that often costs $0.00).

The workflow recipes turn isolated skills into end-to-end chains. workflows.json defines 11 named recipes that run several skills in sequence, passing each output forward. "ship-a-feature" chains ambiguity-resolver to prd-template to rice-prioritisation to roadmap-narrative to go-to-market. "close-the-quarter" chains metrics-framework to churn-analysis to executive-update. Each recipe is exposed over MCP (list_workflows and get_workflow) and as a Claude Code slash command. This is the difference between a pile of skills and a usable professional process.

It is listed in Anthropic's official Claude plugin directory, and the claim is structurally verifiable. The repo ships .claude-plugin/marketplace.json conforming to the official Anthropic marketplace schema (https://anthropic.com/claude-code/marketplace.schema.json), defining 66 plugin bundles (pm-accounting, pm-ai, pm-analytics, pm-career, pm-compliance, pm-design, pm-engineering, pm-finance, pm-legal, pm-security, and 56 more) across 27 professions. The README carries a "Published" badge and explicit instructions to install via /plugin inside Claude Code. This is a semi-official endorsement that most community skill packs do not have.

The Bad

The skill count is inconsistent across every surface, and the numbers are all wrong. The GitHub repository description says "400 professional Agent Skills." The package.json description field says "411 professional Agent Skills." The .claude-plugin/marketplace.json says "411 Claude Skills across 59 bundles covering 27 professions." The README body says "466 professional skills across 65 bundles covering 25 professions." The actual skills/ directory contains 466 skill folders, each with a valid SKILL.md. So the filesystem is right and every prose surface is stale: the README undercounts bundles (65 vs the 66 in plugins/), the package.json undercounts skills by 55, and the GitHub description undercounts by 66. A user reading the README before installing sees a smaller catalog than they actually get, which is a documentation-hygiene failure at scale.

Most skills are doc-only methodology prompts, and their output quality is model-dependent. Of 466 skills, 434 are pure markdown instruction files with no executable component. The 32 that ship calculators are the exception. A skill like prd-template is a well-structured prompt that tells the model how to write a PRD, but whether the resulting PRD is good depends entirely on the model running it. The repo's own eval harness exists to measure this, but the published leaderboard uses Sonnet as both runner and judge by default, so the scores reflect one model's self-consistency more than cross-model robustness. A user who installs this expecting deterministic tools will find mostly prompts.

The README is 110KB and front-loads marketing over orientation. The first scroll is badges, a "Firm" and "Boardroom" multi-agent arena feature, and sponsor appeals before the install command appears. A developer who lands on the repo wanting to know "what is this and how do I install it" has to scroll past the firm, the boardroom, the brain, the context system, and several callouts before reaching the npx pm-claude-skills add line. The install command should be above the fold. The SKILLS.md index (77KB) and CHANGELOG.md (94KB) are thorough but also very large, which makes targeted lookup slow without the search tool the MCP server provides.

The web smoke test hangs and the scripts smoke test depends on a timeout binary not present on macOS by default. Running node tests/web-smoke.mjs hangs past 60 seconds (it appears to start a local server and wait). Running node tests/scripts-smoke.mjs under timeout fails because macOS ships no timeout command without coreutils. These are minor CI-ergonomics issues, but a contributor on macOS running the test suite for the first time will hit both. The repo does not document these prerequisites in the CONTRIBUTING flow.

Smoke Test Results

Tested on a macOS host (aarch64) with Node v24.13.1 and Python 3.12. The repo was shallow-cloned to 176MB (7,686 files). No npm install was needed because the CLI and MCP server are pure Node standard library. Tests covered the validator, the installer (dry-run and real), the MCP server (live JSON-RPC), and two executable skill scripts.

Run A. Fresh clone, skillcheck validator

$ git clone --depth 1 https://github.com/mohitagw15856/pm-claude-skills.git
Cloning into 'pm-claude-skills'... (176MB, 7686 files)

$ node scripts/skillcheck.mjs
 ⚠ assumption-bounty: Description has no "Use when …" trigger clause.
 ⚠ code-simplification: Description has no "Use when …" trigger clause.
 ⚠ verification-before-completion: Description has no "Use when …" trigger clause.

SkillCheck: 466 skills · 0 error(s) · 3 warning(s)
All skills valid. ✓

Pass rate: 466 of 466. Zero structural errors across the entire catalog. The 3 warnings are cosmetic (missing "Use when" trigger phrasing in 3 descriptions out of 466).

Run B. Installer, dry-run and real install

$ node bin/cli.mjs add --agent hermes --dry-run
[dry-run] Installing for 'hermes' into /Users/openclaw/.hermes/skills
 would install 360-feedback-template -> .../360-feedback-template
 would install ab-test-planner -> .../ab-test-planner
 ... (463 more)

$ node bin/cli.mjs add --agent cursor --dry-run
[dry-run] Installing for 'cursor' into ./.cursor/rules
 would install bookkeeping-categorization.mdc -> .../bookkeeping-categorization.mdc
 ... (generated rule files)

$ node bin/cli.mjs add --agent hermes --target /tmp/pmcs-install-test/hermes/skills
Installing for 'hermes' into /tmp/pmcs-install-test/hermes/skills
Installed 466 item(s) for 'hermes'.
Restart hermes: it auto-discovers SKILL.md skills ... by their description. ✓

$ ls /tmp/pmcs-install-test/hermes/skills/ | wc -l
466

$ head -5 /tmp/pmcs-install-test/hermes/skills/ab-test-planner/SKILL.md
name: ab-test-planner
description: "Design statistically rigorous A/B tests..."
--- ✓

Pass rate: 3 of 3. Dry-run previews both native and rule-file installs. Real install copies all 466 skills with frontmatter, references, and templates intact.

Run C. MCP server and executable scripts (functional verification)

$ cat req.jsonl | node mcp/server.mjs
{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2025-10-17",...}} ✓ initialize
{"jsonrpc":"2.0","id":2,"result":{"tools":[run_skill,list_skills,search_skills,
 get_skill,list_workflows,get_workflow]}} ✓ 6 tools
{"jsonrpc":"2.0","id":3,"result":{"content":[{"type":"text","text":
 "- ab-test-planner [production] - Design statistically rigorous A/B tests...
 - ab-test-readout [stable] - Analyse a finished A/B test...
 - experiment-designer [stable] - ..."}],"structuredContent":{...}}} ✓ search
[stderr] MCP server ready: 466 skills, 11 workflow recipes, 6 tools.

$ python3 skills/sprint-planning/scripts/capacity_calculator.py \
 --team 5 --days 10 --velocity 30 --availability 0.8 --carryover 5
Sprint Capacity Estimate
Available focus hours : 240.0
Velocity adj. for availability : 24.0
RECOMMENDED commitment (pts) : 19.2
Capacity for NEW work (pts) : 14.2 ✓

$ npm view pm-claude-skills version dist-tags.latest license
version = '46.0.0'
dist-tags.latest = '46.0.0'
license = 'MIT' ✓

Functional pass rate: 4 of 4 exercised. The MCP server completes the initialize handshake, exposes all six tools with structured output schemas, and returns ranked search results with tier tags. The capacity calculator produces a correct, computed sprint commitment. The npm package is published and live at the claimed version. What could not be exercised: per-skill LLM output quality (model-dependent, covered by the repo's own eval harness which needs an API key).

What the runs tell you

The delivery layer is verifiable end to end without any API key or LLM call. The validator, the installer, the MCP server, and the executable calculators all work on a fresh clone with only Node and Python installed. The honest limit is that the core product (434 doc-only skills) is a set of prompts whose quality depends on the model that applies them, so the eval harness is the right tool for that question rather than a smoke test. The 32 executable skills are the exception: they produce deterministic, checkable numbers.

Setup Walkthrough

1. Install Node 18 or newer (Node 24 used in testing). No Python is required unless you want to run the 32 calculator scripts (Python 3.9-plus, no pip installs needed).
2. Install all 466 skills into Claude, Hermes, Codex, or OpenClaw with one command: npx pm-claude-skills add --agent claude (substitute hermes, codex, or openclaw). No npm install step is needed. The CLI copies skill folders into the right home directory (~/.claude/skills, ~/.hermes/skills, and so on).
3. For Cursor, Windsurf, or Aider, use the same command with a different agent flag: npx pm-claude-skills add --agent cursor. This generates rule files (.mdc for Cursor) in the current project's rules directory.
4. To use it as an MCP server in Claude Desktop or any MCP client, add pm-claude-skills-mcp to your client config: { "mcpServers": { "pm-claude-skills": { "command": "npx", "args": ["-y", "pm-claude-skills-mcp"] } } }. This exposes the search and retrieval tools without installing any files.
5. To install from inside Claude Code, run /plugin and find pm-claude-skills in the marketplace (it is published in Anthropic's official directory).

Post-install gotcha: the installer copies all 466 skills at once, which is a large prompt-context surface if your agent loads every skill description. If you want a subset, install from source and copy only the skill folders you need, or use the MCP server (which retrieves skills on demand via search rather than loading them all). The --link flag symlinks instead of copying on native agents, so updates to the cloned repo propagate without reinstalling.

Alternatives

1. anthropics/skills - the official Anthropic skill pack. Prefer it if you want a small, vendor-curated, first-party set of skills with no third-party author. Prefer pm-claude-skills if you want depth (466 vs dozens) and multi-agent portability.
2. addyosmani/agent-skills - a 77,000-star engineering-focused skill pack. Prefer it if your work is software-engineering-heavy (code review, refactoring, testing patterns). Prefer pm-claude-skills if you need product, go-to-market, compliance, and customer-success coverage alongside engineering.
3. FrancyJGLisboa/agent-skill-creator - a meta-skill for authoring and distributing skills across 17 platforms. Prefer it if your goal is to create your own skills. Prefer pm-claude-skills if your goal is to consume a large ready-made library.
4. legeling/PromptHub - a Prompt/Skill/Agent management toolbox with cloud sync. Prefer it if you want a GUI for managing and syncing skills across machines. Prefer pm-claude-skills if you want the skills themselves plus a CLI installer and MCP server with no account or cloud dependency.